摘要：基础视觉模型能够提供强大的通用表征，但较高的类别级检索准确率并不必然意味着嵌入空间遵循目标语义分类层级。我们在冻结的 DINOv2 特征上研究严格的显式分类层级图像检索，并提出问题：当层级检索效果改善时，这种变化有多少与分类层级感知监督的组织方式有关，又有多少源于欧氏—双曲几何的选择？
我们采用严格的跨类别标准评估更高层级，该标准会排除更细粒度的匹配；并比较使用分类层级距离回归或分类层级感知监督对比目标训练得到的欧氏与双曲投影。一个计算量匹配的 2×2“几何 × 损失”析因实验采用相同的 768-256-32 投影器容量、优化日程、批次顺序以及固定的 100 个 epoch 预算；其中，损失轴表示“分类层级回归（Regression）—Taxonomy-SupCon”目标函数族的对比。在 CUB 上，平均层级 mAP（严格中层/高层平均值，不含 Class/Leaf）中，目标函数族对比在欧氏空间为 +0.0487、在双曲空间为 +0.0414，而几何对比分别仅为 +0.0102 和 +0.0030。在 NABirds 的 Parent-disjoint 检索上，对应的目标函数族对比分别为 +0.0467 和 +0.0440，而几何对比分别为 +0.0017 和 -0.0009。
一项语义对齐控制实验显示，真实分类层级显著优于一个保持结构但随机打乱的层级；与此同时，NABirds 的曲率/半径控制实验并不支持“更强的负曲率”能够解释观察到的层级收益。在两个分类体系中，Regression-to-Taxonomy-SupCon 的对比在总体上大于所评估的几何对比；语义对齐本身也独立地发挥作用，而几何的影响仍然依赖于具体层级。

Abstract:Foundation vision models provide strong generic representations, yet high class-level retrieval accuracy does not necessarily imply that an embedding respects a target semantic taxonomy. We study strict explicit-taxonomy image retrieval on frozen DINOv2 features and ask: when hierarchical retrieval improves, how much of the change is associated with the organization of taxonomy-aware supervision, and how much with the Euclidean-hyperbolic geometry choice?
We evaluate higher levels with strict cross-class criteria that exclude finer-grained matches, and compare Euclidean and hyperbolic projections trained with taxonomy-distance regression or a taxonomy-aware supervised contrastive objective. A compute-matched 2 x 2 Geometry x Loss factorial uses the same 768-256-32 projector capacity, optimization schedule, batch order, and fixed 100-epoch budget; the Loss axis denotes the Regression-to-Taxonomy-SupCon objective-family contrast. On CUB, the objective-family contrasts in mean hierarchy mAP (strict middle/high average, excluding Class/Leaf) are +0.0487 in Euclidean space and +0.0414 in hyperbolic space, compared with geometry contrasts of +0.0102 and +0.0030. On NABirds Parent-disjoint retrieval, the corresponding objective-family contrasts are +0.0467 and +0.0440, whereas geometry contrasts are +0.0017 and -0.0009.
A semantic-alignment control shows that the true taxonomy substantially outperforms a structure-preserving shuffled hierarchy, while a NABirds curvature/radius control does not support stronger negative curvature as the explanation for the observed hierarchy gains. Across the two taxonomies, the Regression-to-Taxonomy-SupCon contrasts are larger in aggregate than the evaluated geometry contrasts; semantic alignment also matters separately, while geometry remains hierarchy-dependent.